Real-time fraud detection has become a critical component in ensuring the security and stability of the digital financial ecosystem. However, existing methods struggle to adapt to the highly dynamic and adversarial nature of modern financial fraud, where malicious actors constantly evolve their strategies to evade detection. To address the dual challenges of complex topological relationships and severe concept drift, we propose the Augmented-Driven Graph Network with Adaptive Exploration (AGNAE). First, this paper introduces an augmented graph neural network tailored for financial transaction graphs, which dynamically models the heterogeneous interactions between transacting entities to capture complex, hidden fraud rings. Second, rather than relying on static classifiers, we rigorously formulate the real-time detection process as a sequential decision-making problem. This paper introduces a deep reinforcement learning agent equipped with an adaptive exploration mechanism to continuously update detection strategies, striking an optimal balance between exploiting known fraud patterns and exploring emerging mutations. Furthermore, a novel joint loss function is designed to synergize topological representation learning with the agent’s long-term financial reward optimization. Extensive experiments on the real-world for IEEE-CIS and FDAD-20 datasets demonstrate that AGNAE significantly outperforms state-of-the-art baselines. Crucially, despite its sophisticated architecture, AGNAE maintains an inference latency of 1.12 ms per transaction, fully satisfying the stringent computational requirements of real-world financial infrastructures.
求 笠木 (2026) studied this question.